clm_8c24d4c24e6e4839b28adf1deff8baa9_r1

2025-06-13_amd-advancing-ai-mi350x-and-mi400-ualoe72-mi500-ual256::AMD25-0012

① SA Source

• Source: 開啟完整 SA 文章
• Section: MI350X and MI355X Specs
• Line hint: 77

Context Before

In late 2027, AMD will release the MI500 UAL256 which will feature 256 physical/logical chips and not just 144 physical/logical chips like in VR300 NVL576

MI350X and MI355X Specs

Evidence

Even though the MI355X uses 1.4x more power, the on-paper specs show that it is less than 10% faster than the MI350X in terms of TFLOPS throughput.

Context After

The on-paper specs for the MI350X and the MI355X are both competitive to the HGX B200 for BF16/FP8/FP4 data types (dtypes). We expect that BF16 and FP8 will be used for training while FP8/FP6/FP4 will be used for inference. On the HGX B200, FP6 shares the same physical circuits as FP8, leading to the same FP8/FP6 on paper FLOP/s. On the MI355X, FP6 shares the same physical circuits as FP4, and so FP6 will have the same peak TFLOP/s speed as FP4. This means that MI355X FP6 is 2.2x faster than B200 FP6. In practice, MI355X FP6 will be at least 20% slower than MI355X FP4 due AI chips always being limited by power.

SemiAnalysis benchmarking has shown ↗ that even though the MI300X and the H100 each show the same on-paper TFLOP/s for FP16 as for BF16 (i.e. Nvidia’s FP16 TF = BF16 = 989 TFLOP/s, AMD’s FP16 = BF16 = 1307 TFLOP/s) in practice – each card delivers different realized TFLOPs when running FP16 vs BF16. We will be publishing an article in the near future running microbenchmarks to figure out a realistic TFLOP/s for MI355X FP6 versus FP4.

② Atomic Claim

MI355X 功耗約為 MI350X 的 1.4 倍，但 paper TFLOPS 提升不到 10%。

• Epistemic Mode: ASSERTED
• Mapping Status: COMPLETE

③ Semantic Frame

{
  "comparison_expression": "1.4x power; <10% more paper TFLOPS",
  "entities": [
    {
      "id": "04_knowledge_base/MI355X",
      "label": "MI355X"
    },
    {
      "id": "04_knowledge_base/MI350X",
      "label": "MI350X"
    },
    {
      "id": "04_knowledge_base/TFLOPS",
      "label": "TFLOPS"
    }
  ],
  "frame_type": "COMPARISON",
  "metric": "power_vs_paper_throughput",
  "operator": "1_4X_POWER_LT_10_PERCENT_SPEED",
  "qualifiers": {
    "condition_text": null,
    "numeric_mentions": [
      "1.4x",
      "10%"
    ],
    "temporal_mentions": []
  }
}

④ Canonical Entity Mapping

Role	Surface Label	Canonical Target
comparison_entity_0	MI355X	MI355X
comparison_entity_1	MI350X	MI350X
comparison_entity_2	TFLOPS	TFLOPS

⑤ Human Review

請在 Properties 逐項確認：

• 原文 → Atomic Claim 是否忠實
• Atomic Claim → Semantic Frame 是否忠實
• Canonical Entity mapping 是否正確
• Epistemic mode 是否保留原文語氣
• 最後選擇 review_action

Review state

Markdown 內文不是正式 approval。只有 Apply bridge 寫入的 Decision Ledger event 才是正式決策。